Qdrant

Qdrant is an open-source, Rust-built vector database for semantic search, retrieval-augmented generation and recommendation workloads at production scale. It combines dense, sparse and multi-vector retrieval with payload filtering applied during index traversal, and ships as a self-hosted binary, a managed cloud on AWS, GCP or Azure, a customer-run hybrid data plane, or an air-gapped private deployment.

Qdrant is an open-source vector search engine written entirely in Rust and licensed under Apache 2.0, built from first principles for vector retrieval rather than bolted onto a general-purpose database. It stores points — vectors carrying arbitrary JSON payloads — and serves approximate nearest-neighbour search over them through REST and gRPC APIs, with official Python and JavaScript clients and a browser UI for inspecting collections. Its distinguishing engineering choice is one-stage filtering: payload conditions are evaluated during HNSW graph traversal rather than before or after it, which keeps latency predictable on the heavily filtered queries that production RAG and recommendation systems actually issue. The engine handles dense, sparse and multi-vector search in a single query, giving native hybrid retrieval with BM25, SPLADE++ and miniCOIL support plus reranking via MMR and late-interaction models. Because memory dominates the cost of vector search, Qdrant ships scalar, binary and asymmetric quantization — the vendor claims up to 64x memory reduction — alongside a custom storage engine called Gridstore and real-time indexing that avoids full rebuilds; version 1.19 added the TurboQuant datatype and memory-tier optimisation. Distribution spans the full range: self-hosted Docker or Kubernetes, Qdrant Cloud managed on AWS, GCP and Azure, a Hybrid Cloud mode that decouples the control plane from a customer-operated data plane, an air-gapped Private Cloud, and Qdrant Edge in beta for on-device retrieval, with serverless announced but not yet shipped. The project reports more than 250 million package downloads and roughly 34,000 GitHub stars, and names Canva, HubSpot, Roche, Bosch, OpenTable, TripAdvisor and Deutsche Telekom among its users. It raised a $50M Series B led by Advance Venture Partners in March 2026, bringing total funding to $87.8M, and competes directly with Pinecone, Weaviate, Milvus and — increasingly — pgvector running inside an existing Postgres estate.

Category

Data & Analytics

Pricing

Usage-based, Freemium, Contact for pricing, Free

Target Market

CTOs, Platform Engineers, ML Engineers, Data Scientists, Enterprise Developers

Deployment

Open-source, Self-hosted, Cloud-first, Hybrid, Multi-cloud, Edge-first

Headquarters

Berlin, Germany

• ✓
  One-stage filtered HNSW search

  Applies payload filters during graph traversal instead of pre- or post-filtering, which avoids the recall collapse and latency spikes that filtered queries cause in naive vector indexes.

• ✓
  Native hybrid and multi-vector retrieval

  Dense, sparse and multi-vector representations coexist in one collection and one query, supporting BM25, SPLADE++ and miniCOIL alongside embeddings without bolting on a second search system.

• ✓
  Quantization and Gridstore engine

  Scalar, binary and asymmetric quantization backed by a purpose-built storage engine cut the RAM footprint that normally dominates vector database cost, with a claimed reduction of up to 64x.

• ✓
  Distributed sharding and replication

  Horizontal sharding with replicas and zero-downtime upgrades lets a single cluster grow past the memory ceiling of one node into the billions of vectors.

• ✓
  Hybrid and Private Cloud deployment

  A Kubernetes-based hybrid mode splits Qdrant's managed control plane from a data plane that never leaves customer infrastructure, and Private Cloud supports fully air-gapped installs.

• ✓
  Reranking and late interaction

  Built-in MMR and late-interaction reranking refine the candidate set inside the database, removing a network hop and a separate reranker service from the retrieval path.

• ✓
  Real-time indexing with REST and gRPC APIs

  New vectors become searchable without a full index rebuild, and both REST and gRPC interfaces ship with official Python and JavaScript clients plus a web UI.

• •
  Permission-aware enterprise RAG

  Serve retrieval for an internal assistant where every query must be filtered by tenant, team and document ACL before results are returned to the model.

• •
  Product and catalogue semantic search

  Replace keyword-only search with hybrid dense plus sparse retrieval so shoppers find items by description and intent, not exact token matches.

• •
  Agent long-term memory

  Store conversation and tool-output embeddings so autonomous agents can recall prior context in sub-second lookups on every step of a run.

• •
  Content recommendation and deduplication

  Compute nearest-neighbour similarity across millions of articles or listings to power related-item rails and to detect near-duplicate submissions at ingest.

• •
  On-device and edge retrieval

  Run the beta Qdrant Edge build on resource-constrained hardware so retrieval works without a round trip to a central cluster.

The engine itself is Apache-2.0 and free to self-host, so the only licence cost is Qdrant Cloud. Cloud billing is metered hourly on actual resource consumption — vCPU, memory, storage, backup storage and paid inference tokens — rather than per seat, and Qdrant does not publish a dollar figure for the Standard tier; you size a cluster and pay for what it consumes. A free forever tier gives a single 0.5 vCPU / 1GB node for prototypes. Premium requires a minimum spend and is where SSO, private VPC links, multi-AZ topology, customer-managed encryption keys, the 99.9% SLA and 24/7 support live. Hybrid Cloud and Private Cloud are sales-quoted. Practical budgeting note: memory is the dominant cost driver in vector search, which is why the quantization features are a pricing feature as much as a performance one.

This page was written from 7 sources, 4 on domains other than qdrant.tech.

1. 1.qdrant.tech — qdrant.techvendor
2. 2.qdrant.tech — pricingvendor
3. 3.qdrant.tech — series b announcementvendor
4. 4.github.com — qdrant
5. 5.techtarget.com — Qdrant raises 50M in funding to fuel vector database growth
6. 6.encore.dev — pgvector vs qdrant
7. 7.hn.algolia.com — hn.algolia.com